|Publication number||US7440633 B2|
|Application number||US 10/741,913|
|Publication date||Oct 21, 2008|
|Filing date||Dec 19, 2003|
|Priority date||Dec 19, 2003|
|Also published as||US20050135694|
|Publication number||10741913, 741913, US 7440633 B2, US 7440633B2, US-B2-7440633, US7440633 B2, US7440633B2|
|Inventors||Scott J. Daly|
|Original Assignee||Sharp Laboratories Of America, Inc.|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (46), Non-Patent Citations (1), Classifications (14), Legal Events (4)|
|External Links: USPTO, USPTO Assignment, Espacenet|
This invention relates to a system for image enhancement and, more particularly, to a system for enhancing the quality of a quantized image.
As the state of the art of digital signal technology advances, related technologies such as digital image processing has experienced a corresponding growth and benefit. For example, the development and proliferation of facsimile communications allows images to be encoded into digital signals, transmitted over conventional telephone line, and decoded into a close representation of the original images. Image data are also digitally encoded for ease of storage, modification, copying, etc. As is common experience with growing technologies, the field of digital image processing is also experiencing problems with applications in new areas.
Problems in the area of digital image processing relate generally to achieving a balance between acceptable image distortion and bit-depth representations. In order to increase the efficiency and therefore the usefulness of digital image decoding schemes, the coding system must provide a coded set of image data that is more efficient to store, transmit, etc., than the original image data and must reproduce a decoded image with some minimum level of quality. However, the conversion of relatively high bit rate image data to lower bit rate data virtually always entails a loss of image quality.
One straightforward method for digitizing an image is to create an artificial grid over the image and to assign a value to each grid space representing the color of the original image at that grid space. If the grids are made small enough and the values represent a large enough range of color, then the image may be encoded and decoded with small image quality degradation. For example, display screen images are made up of an array of pixels, i.e., picture elements. On a black and white screen, each pixel has a value of one or zero representing the on/off state of the pixel. In a one-to-one bit-to-pixel coding scheme, each pixel value is represented as a 1 or as a 0 and the entire screen image is encoded. The result of the encoding is an array of binary values. To decode the image, the array values are translated into a screen image having pixels on or off in the same order in which they were originally encoded.
If the image is comprised of more than two distinct colors, then more than a 1-bit code must be used to represent the pixel values. For example, if four distinct colors are present in the image, a 2-bit binary code can represent all of the values. If the image includes 256 distinct colors, an 8-bit binary code is required to uniquely represent each of the color values. The memory requirements for such coding schemes increase as the number of distinct colors in the image increases. However, with high bit-depth representation schemes, the quality of the image that results will be good as long as the digital image transmission or recovery from storage is successful.
To reduce the size of the encoded digital image, the bit-depth representation of the image may be reduced in some manner. For example, an image with a bit-depth of 6 bits per pixel requires significantly less storage capacity and bandwidth for transmission than the same sized image with 16 bits per pixel.
Decoded images, constructed by a low bit-depth representation, generally suffer from the following types of degradations: (a) quasi-constant or slowly varying regions suffer from contouring effects and amplified granular noise, and (b) textured regions lose detail.
Contouring effects, which are the result of spatial variations, in a decoded image are generally fairly obvious to the naked eye. The contouring effects that appear in the slowly varying regions are also caused by the fact that not all of the variations in the intensity of the original image are available for the decoded image. For example, if a region of the original image included an area having 4 intensity changes therein, the decoded image might represent the area with only 2 intensities. In contrast to the contouring effects, the effect of the granular noise on the viewed image is often mitigated by the very nature of the textured regions. But it can be both, amplified or suppressed due to quantization, as well as altered in spectral appearance.
Kundu et al., U.S. Pat. No. 5,218,649, disclose an image processing technique that enhances images by reducing contouring effects. The enhancement system identifies the edge and non-edge regions in the decoded image. Different filters are applied to each of these regions and then they are combined together. A low pass filter (LPF) is used on the non-edge regions, and a high-pass enhancer is used on the edge regions. Kundu et al. teaches that the contour artifacts are most visible in the non-edge areas, and the LPF will remove these edges. Unfortunately, problems arise in properly identifying the edge regions (is a steep slope an edge or a non-edge?). In addition, problems arise in setting thresholds in the segmentation process because if the contours have a high enough amplitude, then they will be classified as edges in the segmentation, and thus not filtered out. Moreover, the image segmentation requires a switch, or if statement, and two full-size image buffers to store the edge map and smooth map, because the size of these regions varies from image to image, all of which is expensive and computationally intensive.
Chan, U.S. Pat. No. 5,651,078, discloses a system that reduces the contouring in images reproduced from compressed video signals. Chan teaches that some contours result from discrete cosine transformation (DCT) compression, which are more commonly called blocking artifacts. Chan observes that this problem is most noticeable in the dark regions of image and as a result adds noise to certain DCT coefficients when the block's DC term is low (corresponding to a dark area). The resulting noise masks the contour artifacts, however, the dark areas become noisier as a result. The dark areas thus have a noise similar to film-grain-like noise which may be preferable to the blocking artifacts.
What is desirable is a system for reducing contouring effects of an image. Because the system does not necessarily affect the encoding or transmission processes, it can be readily integrated into established systems.
The present inventor observed that the many of the objectionable contouring effects in images appear in regions of the image that are generally free from a significant number of edges or otherwise high texture. The present inventor similarly observed fewer objectionable contouring effects in regions of the image that have a significant number of edges or otherwise high texture. After consideration of this difference in observing objectionable contouring effects, the present inventor considered that the human visual system has lower sensitivity to these contour effects in the high frequency regions of the image and the human visual system has higher sensitivity to these contour effects in the low frequency regions of the image.
In order to preserve the quality of the image, the system preferably (albeit not necessary) reduces the effects of the objectionable contours without having to add noise to the image in order to hide them. Thus the system is suitable for use with images that are otherwise free of image capture noise, such as computer graphics and line art with gradients. Moreover, it would be preferable (albeit not necessary) that the technique is implemented without decision steps or if statements to achieve computational efficiency. Also, the technique should require not more than a single buffer of a size less than or equal to the image, and more preferable a buffer less than 30, 20, or 10 percent of the size of the image.
To reduce the false step edges of contouring the image is preferably low-pass filtered 102. The low-pass filter applied to the image also acts to increase the bit depth, since the low pass filter is primarily an averaging operation (averaging across a neighborhood of pixels), and the averaging results in a higher precision. Other techniques may likewise be applied to effectively modify the bit depth of the image. In some cases, the system may also modify the bit depth by switching the number of bits used for processing the image, such as 8 bits to 16 bits. One way of characterizing an increase in the bit depth of an image is to modify in any manner an image to 2N different levels, where N≠P. Alternatively, the image may already have a bit-depth needed/desired for the output (P=N), but the image itself may have a limited bit depth of P (P<N) from a previous operation. In this case, the value of P should be known or otherwise determined. Alternatively, the image may have a bit-depth of N, but the image itself may have a limited bit-depth of P (P<N).
The result of the low pass filtering of the image is to modify the image to achieve a bit depth of N. In most cases the bit depth N is the desired bit depth in the final image, such as an image having bit depth N to be displayed on a N bit display. The low pass filter should be sufficiently wide (in the spatial domain) to reduce most false steps (contouring) due to the bit-depth limit P. It is noted that the low pass filter likewise reduces other desirable image information within the image, so merely low-pass filtering the image is insufficient. Note that the low pass filter also reduces much useful image information, such as by severely blurring the image, so this step is insufficient to properly rectify the undesirable contouring.
The resulting image from the low pass filter is primarily the low frequency components of the image. It is likewise to be understood that any suitable filter may be used, such as one that generally attenuates the high frequency components with respect to the low frequency components of the image.
The resulting image from the low pass filter is then used to predict or otherwise simulate the contouring that would have occurred in the low-frequency aspects of the image. The resulting image from the low pass filter is then quantized to the bit-depth P (e.g., the original image bit-depth) or other bit depth at block 104. Since the image prior to being quantized has reduced edge sharpness and reduced high frequency content, the edges that occur as a result of quantizing the image are representative of false contouring in the original image. Accordingly, the result is an estimation of the false contours of the image caused by the quantization process of bit-depth P, especially those false contours in the low frequency regions of the image which are objectionable to the viewer.
After estimating the false contouring of the image, the system may subtract the image containing estimates of the false contours from the original bit-depth limited image 106. This in essence reduces the low-frequency content of the input image, while primarily leaving the high frequency content. However, the subtraction also results in another important attribute, namely, it also removes the high frequencies that are a result of the quantization appearing in the low frequency regions of the image. That is, the high frequency and low frequency components of the contour artifacts are reduced to the extent that the estimated contour artifacts are correct.
It would be noted that the system does not significantly remove the quantization artifacts in the higher frequency aspects of the image. This is because, while the contours estimated by the quantization are fairly accurate in low frequency components, that are not correct in high frequency components because it starts with a merely low pass filtered image issued at step 104. As previously noted, there is limited need to remove the quantization artifacts in the higher frequency aspects of the image because they are masked by the actual high frequency image content. For purposes of identification, this resulting image may be referred to as the high pass image 108. The subtraction process leads to a bit-depth increase of 1 due to the sign. For example a subtraction of two images with 0 to +255 results in an image with a range from −255 to +255. Hence the high pass image 108 has N+1 bit depth (this is based upon a source image being padded to N bits before the subtraction operation). Padding may include inserting 0's and/or 1's to the least significant bits that are missing from the lower bit depth representations relative to the higher.
It is noted that the output of the filter may not need N+1 bits, but it should be able to carry a sign, and thus have a zero mean level. If only N bits are used and one bit is dedicated to the sign, then only N−1 bits are available for addition to the low pass filter image (in the last addition step) at 100. In that case some edge sharpness and detail may be lost.
The next step is to add the high pass image 108 and the filtered low pass components 112 together at 110. It is preferable to use the low pass filtered version of the source image for the low pass portion 110, as opposed to the quantized version of the low pass filtered image because the low pass filtering acts to create a ramp across the steps of the false contours. The result of this addition operation is N+1 bits. It is N+1 since the range may be larger than the input image (for example if the low pass component for N bits may be 0 to +255, the high pass component may have a range of N+1 (−255 to +255) and the result is −255 to +512). Anything out of the range of N bits is clipped (e.g., out of range 0 to +255 for N=8). It turns out that there are a limited number of pixels that fall out of that range, and when they do, they are usually isolated edge pixels. The clipping that occurs as a result is not readily visible in the final image 114.
For purposes of illustration, example waveforms associated with these primarily image modifications are shown in
As shown in
The two variations illustrated in
The number of gray levels is given by N or P, which are bit-depths, so the numbers of levels is 2N or 2P. However, another embodiment is to not use bit-depth to determine the number of gray levels, but the numbers of gray levels directly.
In a particular implementation, each of the steps shown in
The preferred embodiment uses a version where the low pass filter is implemented in two Cartesian separable steps (i.e., a cascade of horizontal and vertical steps), and the filter width spatially is 10 pixels, with a uniform impulse response. A two-dimensional filter may likewise be used, such as a two-dimensional Gaussian filter.
In a typical implementation the bit-depth of the input and output images are known in advance. This is typically the case in many applications, such as display an eight bit image on a 10-bit display.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US4020462||Dec 8, 1975||Apr 26, 1977||International Business Machines Corporation||Method and apparatus for form removal from contour compressed image data|
|US4097845 *||Aug 18, 1977||Jun 27, 1978||Rush-Presbyterian-St. Luke's Medical Center||Method of and an apparatus for automatic classification of red blood cells|
|US4196452||Dec 1, 1978||Apr 1, 1980||Xerox Corporation||Tone error control for image contour removal|
|US4223340||May 11, 1979||Sep 16, 1980||Rca Corporation||Image detail improvement in a vertical detail enhancement system|
|US4268864||Dec 5, 1979||May 19, 1981||Cbs Inc.||Image enhancement system for television|
|US4399461||Mar 12, 1982||Aug 16, 1983||Eastman Kodak Company||Electronic image processing|
|US4402006||Feb 23, 1981||Aug 30, 1983||Karlock James A||Image enhancer apparatus|
|US4523230||Oct 22, 1984||Jun 11, 1985||Rca Corporation||System for coring an image-representing signal|
|US4536796||Aug 23, 1983||Aug 20, 1985||Rca Corporation||Non-linear dynamic coring circuit for video signals|
|US4549212||Aug 11, 1983||Oct 22, 1985||Eastman Kodak Company||Image processing method using a collapsed Walsh-Hadamard transform|
|US4553165||Aug 11, 1983||Nov 12, 1985||Eastman Kodak Company||Transform processing method for reducing noise in an image|
|US4613986 *||Jun 14, 1984||Sep 23, 1986||Ncr Canada Ltd.||Digital filter for enhancement of document images|
|US4962426||Apr 7, 1989||Oct 9, 1990||Hitachi, Ltd.||Dynamic noise reduction circuit for image luminance signal|
|US4977605||Sep 28, 1987||Dec 11, 1990||Bertin & Cie||Binary quantification of an image having multiple levels of greys|
|US5025312||Mar 30, 1990||Jun 18, 1991||Faroudja Y C||Motion-adaptive video noise reduction system using recirculation and coring|
|US5218649||May 4, 1990||Jun 8, 1993||U S West Advanced Technologies, Inc.||Image enhancement system|
|US5227869||Aug 15, 1991||Jul 13, 1993||Ikegami Tsushinki Co., Ltd.||Method for correcting contour of image|
|US5260791||Jun 4, 1992||Nov 9, 1993||David Sarnoff Research Center, Inc.||Method and apparatus for the spatio-temporal coring of images|
|US5389978||Mar 1, 1993||Feb 14, 1995||Samsung Electronics Co., Ltd.||Noise eliminative circuit employing a coring circuit|
|US5420431 *||Jul 26, 1993||May 30, 1995||Kabushiki Kaisha Toshiba||Imaging apparatus using multiple image signal read out lines with improved image quality|
|US5438633 *||Mar 9, 1993||Aug 1, 1995||Eastman Kodak Company||Method and apparatus for gray-level quantization|
|US5515180 *||Nov 16, 1993||May 7, 1996||Sharp Kabushiki Kaisha||Image processing device|
|US5526446||Sep 24, 1991||Jun 11, 1996||Massachusetts Institute Of Technology||Noise reduction system|
|US5539866 *||May 11, 1994||Jul 23, 1996||Xerox Corporation||Method and apparatus for accurately rendering half-bitted image pixels|
|US5585818 *||May 14, 1993||Dec 17, 1996||Canon Kabushiki Kaisha||Display control unit and display control method|
|US5587711 *||Sep 30, 1994||Dec 24, 1996||Apple Computer, Inc.||Method and system for reconstructing quantized signals|
|US5651078 *||Jul 18, 1994||Jul 22, 1997||Thomson Consumer Electronics, Inc.||Method and apparatus for reducing contouring in video compression|
|US5696852||Mar 31, 1995||Dec 9, 1997||Canon Kabushiki Kaisha||Image signal processing apparatus|
|US5844611 *||May 30, 1995||Dec 1, 1998||Fujitsu Limited||Image coding system which limits number of variable length code words|
|US5912992||Mar 21, 1997||Jun 15, 1999||Sharp Kabushiki Kaisha||Binary image forming device with shading correction means using interpolation of shade densities determined by using sample points|
|US5920653||Oct 22, 1996||Jul 6, 1999||Hewlett-Packard Company||Multiple spatial channel printing|
|US5956470 *||Sep 25, 1997||Sep 21, 1999||Xerox Corporation||Text quality enhancement via resolution enhancement technique based on separating jaggedness detection and filtering|
|US6038576 *||Dec 2, 1997||Mar 14, 2000||Digital Equipment Corporation||Bit-depth increase by bit replication|
|US6055340||Feb 27, 1998||Apr 25, 2000||Fuji Photo Film Co., Ltd.||Method and apparatus for processing digital images to suppress their noise and enhancing their sharpness|
|US6381373 *||Aug 31, 1999||Apr 30, 2002||Sony Corporation||Apparatus and method for reducing quantization error in digital image signals|
|US6441867 *||Oct 22, 1999||Aug 27, 2002||Sharp Laboratories Of America, Incorporated||Bit-depth extension of digital displays using noise|
|US6504953||Aug 18, 1999||Jan 7, 2003||Heidelberger Druckmaschinen Aktiengesellschaft||Method for the automatic removal of image errors|
|US6522784 *||Apr 11, 2000||Feb 18, 2003||International Business Machines Corporation||Enhanced compression of gray-level images|
|US6546143 *||Mar 12, 1999||Apr 8, 2003||Hewlett-Packard Development Company||Efficient wavelet-based compression of large images|
|US6642962 *||Sep 1, 1999||Nov 4, 2003||Neomagic Corp.||Merged pipeline for color interpolation and edge enhancement of digital images|
|US6661421 *||Mar 8, 1999||Dec 9, 2003||Mitsubishi Electric & Electronics Usa, Inc.||Methods for operation of semiconductor memory|
|US20010031083 *||Nov 30, 2000||Oct 18, 2001||Eastman Kodak Company||Method of image processing|
|US20040252201 *||Jun 10, 2003||Dec 16, 2004||Zoran Corporation||Digital camera with reduced image buffer memory and minimal processing for recycling through a service center|
|JPH089154A||Title not available|
|JPH03102579A||Title not available|
|JPH09261464A||Title not available|
|1||Tilne Jarske, Petri Haavisto and Irek Defee, "Post-Filtering Methods for Reducing Blocking Effects From coded Images," IEEE, Jun. 27, 1994.|
|U.S. Classification||382/254, 382/100|
|International Classification||G06K9/36, G06T5/00|
|Cooperative Classification||H04N19/86, H04N19/44, H04N19/126, G06T5/002, G06T5/20|
|European Classification||G06T5/00D, H04N7/26P4, H04N7/26D, H04N7/26A4Q2, H04N7/30E4|
|Apr 26, 2004||AS||Assignment|
Owner name: SHARP LABORATORIES OF AMERICA, INC., MASSACHUSETTS
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:DALY, SCOTT J.;REEL/FRAME:015255/0009
Effective date: 20040419
|Nov 12, 2008||AS||Assignment|
Owner name: SHARP KABUSHIKI KAISHA, JAPAN
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:SHARP LABORATORIES OF AMERICA INC.;REEL/FRAME:021817/0899
Effective date: 20081112
|Apr 3, 2012||FPAY||Fee payment|
Year of fee payment: 4
|Apr 13, 2016||FPAY||Fee payment|
Year of fee payment: 8